Token Level Identification of Multiword Expressions Using Contextual Information

Studies on detecting idiomatic expressions mostly focus on discovering potentially idiomatic expressions disregarding the context. However, many idioms like kick the bucket could be idiomatic/literal depending on the context. In this work, we use Context2Vec model to include contextual information. The model learns a generic context embedding function from large corpora, using bidirectional LSTM. We build a simple nearest neighbor classification on Context2Vec which outperforms the popular context representation of average-of-word-embeddings. Through lexical substitution task, we further show that the Context2Vec model is able to place MWEs into distinct {`}sense{'}(idiomatic/literal) regions of the embedding space, while traditional word embedding i.e. Skip Gram lacks this ability.